75 lines
2.7 KiB
Docker
75 lines
2.7 KiB
Docker
############################
|
||
# Stage-0: 取运行依赖轮子 #
|
||
############################
|
||
ARG CUDA_VERSION=12.8.1 # 仅 12.6.1 / 12.8.1 经测试可用
|
||
FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu22.04 AS builder
|
||
|
||
# ---- Python & pip ----
|
||
RUN apt-get update && \
|
||
apt-get install -y --no-install-recommends python3 python3-pip && \
|
||
ln -sf /usr/bin/python3 /usr/bin/python && \
|
||
python -m pip install --no-cache-dir --upgrade pip
|
||
|
||
# ---- 安装 PyTorch + SGLang ----
|
||
ARG TORCH_VER=2.3.0 # 如需别的版本自行改
|
||
RUN case "$CUDA_VERSION" in \
|
||
12.6.1) CUINDEX=126 ;; \
|
||
12.8.1) CUINDEX=128 ;; \
|
||
*) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
|
||
esac && \
|
||
python -m pip install --no-cache-dir \
|
||
torch==${TORCH_VER}+cu${CUINDEX} \
|
||
torchvision==0.18.0+cu${CUINDEX} \
|
||
--extra-index-url https://download.pytorch.org/whl/cu${CUINDEX} && \
|
||
python -m pip install --no-cache-dir sglang==0.4.8.post1 && \
|
||
# 12.8 额外装官方 sgl_kernel & NCCL
|
||
if [ "$CUDA_VERSION" = "12.8.1" ]; then \
|
||
python -m pip install --no-cache-dir nvidia-nccl-cu12==2.27.3 --force-reinstall --no-deps && \
|
||
python -m pip install --no-cache-dir \
|
||
https://github.com/sgl-project/whl/releases/download/v0.2.0/sgl_kernel-0.2.0+cu128-cp39-abi3-manylinux2014_x86_64.whl \
|
||
--force-reinstall --no-deps ; \
|
||
fi
|
||
|
||
############################
|
||
# Stage-1: 最小运行镜像 #
|
||
############################
|
||
FROM nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu22.04
|
||
|
||
ENV DEBIAN_FRONTEND=noninteractive \
|
||
PYTHONUNBUFFERED=1
|
||
|
||
# 基础 Python
|
||
RUN apt-get update && \
|
||
apt-get install -y --no-install-recommends python3 python3-distutils && \
|
||
ln -sf /usr/bin/python3 /usr/bin/python && \
|
||
rm -rf /var/lib/apt/lists/*
|
||
|
||
# 复制虚拟环境（即 site-packages）到最终镜像
|
||
COPY --from=builder /usr/local/lib/python3.*/dist-packages /usr/local/lib/python3.*/dist-packages
|
||
COPY --from=builder /usr/local/bin /usr/local/bin
|
||
|
||
# 拷贝模型文件（示例路径，按需修改）
|
||
COPY ./Alibaba/Qwen3-8B /root/.cradle/Alibaba/Qwen3-8B
|
||
|
||
# 暴露推理端口
|
||
EXPOSE 30000
|
||
|
||
# 默认启动 SGLang 推理服务
|
||
# CMD ["python3", "-m", "sglang.launch_server", \
|
||
# "--cpu-offload-gb", "64", \
|
||
# "--host", "0.0.0.0", \
|
||
# "--port", "30000", \
|
||
# "--model-path", "/root/.cradle/Alibaba/QwQ-32B/", \
|
||
# "--tp", "2", \
|
||
# "--api-key", "token-abc123", \
|
||
# "--mem-fraction-static", "0.8"]
|
||
|
||
|
||
CMD ["python3", "-m", "sglang.launch_server", \
|
||
"--host", "0.0.0.0", \
|
||
"--port", "30000", \
|
||
"--model-path", "/root/.cradle/Alibaba/Qwen3-8B/", \
|
||
"--tp", "1", \
|
||
"--api-key", "token-abc123"]
|
||
|